Applying satellite observations of tropical cyclone internal structures

Applying satellite observations of tropical cyclone internal structures to rapid intensification forecast with machine learning Hui Su 1, Longtao Wu 1, Raksha Pai 2, Alex Liu 3, Peyman Tavallali 1, Albert J. Zhai 4, Jonathan H. Jiang 1, Mark De. Maria 5 1 Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA, 91109 2 IBM Analytics Services 3 RMDS Lab 4 Dept. of Computing and Mathematical Sciences, California Institute of Technology 5 National Hurricane Center, Miami, FL NOAA AI workshop, November 5, 2020

Motivation • Strong winds from hurricanes, and flooding from storm surge and heavy rain, cause losses on various sectors of the economy. • In US, hurricane loss averages $24 billion annually from 1980 to 2019 (in 2019 dollars, inflation adjusted), about 54% of the cost of all major disasters including droughts, wildfires and other severe storms. • National Hurricane Center (NHC) tropical cyclone (TC) intensity forecast skills have not shown much improvement in the last few decades, while the TC track errors have declined steadily. • Storms that undergo rapid intensification (RI, defined as hurricane maximum sustained wind speed change greater than 30 knots within 24 hours) are associated with the highest forecast errors and larger economic losses. About 80% of major hurricanes in the Atlantic basin experienced RI. • The NHC's probability of detection (POD) for RI in the Atlantic basin is < 40% (Kaplan et al. 2015).

Factors Controlling Hurricane Intensification • • • Warm sea surface temperature Weak vertical wind shear Moist environment Strong symmetric convection Intense convective bursts Cold outflow temperature

Hurricane Forecast Models • Numerical models

Statistical Forecast Models Predictors Predictands

Statistical Hurricane Intensity Prediction Scheme (SHIPS) Predictors for RI Kaplan et al. (2015, Weather and Forecasting)

Current NHC RI Forecast Skills Probability Of Detection (POD) False Alarm Ratio (FAR) Peirce Skill Score (PSS) PSS POD FAR Kaplan et al. (2015, Weather and Forecasting)

Path to Improve RI Forecast Predictors Predictands

New Predictors for RI • Our analysis of NASA satellite observations identifies several new predictors for TC intensity change, including inner core precipitation structure, ice water content, and outflow temperature TC Intensity TC Intensification Rate Sample size for TRMM data (1998 -2014)

Surplus Inner-Precipitation Surplus Precipitation Su et al. (2020, GRL)

Cloud. Sat Ice Water Content TC Intensity TC Intensification Rate

MERRA-2 Ice Water Path (IWP) TC Intensification Rate TC Intensity 0 -100 km

Aura MLS Outflow Temperature TC Intensification Rate TC Intensity 0 -200 km

MERRA-2 Precipitable Water Vapor TC Intensification Rate TC Intensity 0 -300 km

MERRA-2 Liquid Water Path TC Intensification (LWP) Rate TC Intensity 0 -100 km • Cloud. Sat LWP does not show similar relation.

Exploration with IBM Watson Studio • We use the IBM Watson Studio to facilitate the machine learning model building and analysis. NASA-JPL IBM Watson Studio: Evaluation of end-to-end ML workflow Prepare Data IBM (in-kind) NOAA-NHC (in-kind) Build ML Models Deploy Models Visualize Results

Training ML Models with the SHIPS and New Predictors o ATL Training: 2680 total cases from 1998 -2008 o EPAC Training: 2428 total cases from 1998 -2008 Kaplan et al. (2015, Weather and Forecasting)

Challenges in Operational Forecast

Improved RI Forecast with ML o ATL o Training: 2680 total cases from 1998 -2008 o Test: 1228 total cases from 2009 -2014 o EPAC o Training: 2428 total cases from 1998 -2008 o Test: 1349 total cases from 2009 -2014

Sensitivity to Different Predictors

Realtime Forecast System https: //tcis. jpl. nasa. gov/data/ships/

Summary • Tropical cyclone intensity change is approximately linearly correlated with surplus inner-core precipitation, ice water path and outflow temperature above a threshold value. • The threshold value can be determined from neutral TCs of nearly constant intensity and it varies approximately linearly with TC intensity. • The JPL-ML model significantly outperforms the NHC operational RI consensus forecast. Our probability of detection for RI in the Atlantic Basin is 40%, 60% and 200% higher than the NHC operational model while the false alarm ratio is only 4%, 7% and 6% higher for 25 -, 30 - and 35 -kt RI thresholds, respectively. Su, H. , Wu, L. , Jiang, J. H. , Pai, R. , Liu, A. , Zhai, A. J. , et al. , Applying satellite observations of tropical cyclone internal structures to rapid intensification forecast with machine learning. Geophysical Research Letters, 47, e 2020 GL 089102, http: //dx. doi. org/10. 1029/2020 GL 089102, 2020. NASA press release: https: //www. jpl. nasa. gov/news. php? feature=7738.